Source fingerprinting loess deposits in Central Asia using elemental geochemistry with Bayesian and GLUE models

The provenance of loess deposits in Central Asia is largely unexplored. Accordingly, the goals of this research
were to test and compare the performance of two different models (generalized likelihood uncertainty estimation
- GLUE and a Bayesian model) for quantifying the uncertainty in source apportionment estimated for 46 target
loess samples collected in the Ili basin, in eastern Central Asia. Model performance was evaluated using goodness-
of-fit (GOF), mean absolute fit (MAF) and virtual mixtures (VM) in combination with root mean square
error (RMSE) and mean absolute error (MAE). Our dataset comprised 132 surficial samples collected from three
potential sources comprising river alluvium (n = 29), sand dunes (n = 35) and topsoils (n = 68). All samples
were analysed for elemental geochemistry. Six geochemical properties (Co, Er, Y, Ga, Dy and Pb) were selected
in a composite fingerprint which classified 83% of the samples from the three source categories correctly. Based
on both models, source contributions to the loess samples were in the following order: topsoils > river alluvium
> sand dunes. Based on the GOF and MAF tests, both models were accurate in predicting measured tracer
concentrations in the loess samples. The Bayesian model was slightly more accurate (mean RMSE 1.6%, mean
MAE 1.8%) than the GLUE (mean RMSE 5.0%, mean MAE 4.7%) model in predicting known source contributions.
Overall, our results provide confirmation that application of source fingerprinting with elemental geochemistry
and uncertainty modelling techniques is useful for identifying the provenance of loess sediments in
arid and desert environments.